The short version
Every resume tailored on JobSearchFix is run through Anthropic's Claude — currently Claude Opus 4.5 for the heavy lifting and Claude 3.5 Sonnet for the fast ATS scoring pass. Not ChatGPT. Not Gemini. Not a local Llama. Claude.
That's not a marketing choice. It's what won the bake-off on real client resumes across our coaching caseload.
What I actually tested
I ran the same 40 resumes — a mix of transitioning military, veterans, cleared professionals, and all job seekers — through GPT-4o, Gemini 1.5/2.0, Llama 3.1 70B, and Claude (Sonnet + Opus). Same job descriptions. Same prompt. Same PII-scrubbed inputs.
I graded on four things that actually matter for a job search:
1. Truthfulness — did it invent employers, dates, metrics, or credentials? 2. Structure discipline — did it respect reverse-chronological order, 2-page cap, and the 10-year civilian lookback (while keeping ALL military experience)? 3. Military translation — did it turn "Platoon Sergeant" into a Fortune-500-legible bullet with scope, budget, and outcome? 4. ATS keyword integration — did it weave JD phrases in naturally, or keyword-stuff them into a skills list?
Where Claude wins
Truthfulness. This is the big one. GPT-4o invented a metric on 11 of 40 resumes ("increased revenue 23%" when no revenue number existed in the source). Gemini invented employer names twice. Claude invented nothing across all 40 — when a number was missing, it inserted a [QUANTIFY: e.g., team size] placeholder for the candidate to fill in. That's exactly the behavior I want in a coach.
Following long, layered instructions. The JobSearchFix system prompt is ~800 words: PII rules, structure rules, military lookback rules, formatting rules, tone rules. Claude follows all of them in one pass. GPT-4o typically forgets 1–2 rules per generation and needs a follow-up. Gemini reorders sections unpredictably.
Prose quality on senior resumes. For director-and-above bullets, Claude writes tighter, more executive-sounding language without the corporate LinkedIn-ese ("passionate," "results-driven," "seasoned"). It sounds like a résumé, not a cover letter.
Refusing to over-embellish. Ask Claude to "make this bullet more impressive" and it asks you for the underlying facts. Ask GPT-4o and it invents the facts.
Where the others were fine
I'm not going to pretend Claude wins everything. GPT-4o is faster on short prompts and cheaper at scale. Gemini has a longer context window. Llama 3.1 is free if you self-host. For chit-chat, all three are perfectly good.
But resume tailoring is not chit-chat. It's a high-stakes, high-precision, high-truthfulness task where a hallucinated metric can cost you the job — or worse, get you fired in the reference check.
The privacy layer
Model choice is only half the story. Before your resume text reaches Claude, it's scrubbed twice:
- In your browser (
src/lib/scrub-pii.ts) — SSNs, emails, US phone numbers stripped. - Again on the server (
_shared/coach.ts) — same scrubbers, plus name and street-address removal. ZIP, city, and state stay because location context matters.
Anthropic's API is not used to train their models on customer data. Combined with our scrub layer, Claude sees a text blob that has no PII in it, tailors your resume, and returns the result. Nothing is retained by the model.
Why this matters to you
If you're pasting your resume into a free ChatGPT tab and asking it to "tailor this for [job]," you are:
1. Sending your PII to a training-eligible endpoint. 2. Using a model that will happily invent metrics to sound more impressive. 3. Getting output that doesn't respect the structure rules a real coach enforces.
That's a bad trade. The JobSearchFix engine exists so you don't have to make it.
Try it
- [Resume Tailoring](/dashboard/resumes): Paste the JD, attach your resume, get a Claude-tailored draft with an ATS score before and after.
- [LinkedIn Optimizer](/dashboard/linkedin): Same engine, applied to your headline, About, and work experience.
If you want to nerd out on the prompt design, ask K.E.N.N.Y. — the chat coach in the corner runs on the same models.
— The JobSearchFix Team